Japan Unveils Next-Generation, Pascal-Based AI Supercomputer (nextplatform.com)

The Tokyo Institute of Technology has announced plans to launch Japan's "fastest AI supercomputer" this summer. The supercomputer is called Tsubame 3.0 and will use Nvidia's latest Pascal-based Tesla P100 GPU accelerators to double its performance over its predecessor, the Tsubame 2.5. Slashdot reader kipperstem77 shares an excerpt from a report via The Next Platform: With all of those CPUs and GPUs, Tsubame 3.0 will have 12.15 petaflops of peak double precision performance, and is rated at 24.3 petaflops single precision and, importantly, is rated at 47.2 petaflops at the half precision that is important for neural networks employed in deep learning applications. When added to the existing Tsubame 2.5 machine and the experimental immersion-cooled Tsubame-KFC system, TiTech will have a total of 6,720 GPUs to bring to bear on workloads, adding up to a total of 64.3 aggregate petaflops at half precision. (This is interesting to us because that means Nvidia has worked with TiTech to get half precision working on Kepler GPUs, which did not formally support half precision.)

Comments

Pascal-based?

[davidwr]

Am I the only one that thought "LISP machines, okay, but Pascal?"

RTFA

[subk]

Pascal, the GPU design. Not Pascal, the language.

GPUs have limited applications

[140Mandak262Jamuna]

It keeps coming back. Massively parallel machine, thousands of cores all working in parallel. Naively multiply add up all the megaflops and get some massive number and tout it big. We can simply add all the flops of all the servers in some Amazon cloud and claim that is the super computer. Back in the 80s "transputer" was all the rage. Before that it was the "vector" computers. Then "the network *is* the computer", then GPUs...

As of now there are very few applications for massively souped up GPU processes. Fluid mechanics loves this GPU. Navier-Stokes is probably the most difficult equation to solve, agreed. But it is hyperbolic, with limited "zone of influence", and numerical equations are quite simple, just mass, momentum and energy balances in the control volume. It plays well in GPU, the calculations fit inside the teensy memory and processor. All time domain problems are hyperbolic and they all can be ported to GPU, theoretically. But try squeezing Maxwell's Equations into that teensy processor!

Graphics card companies are desperately looking for new markets and they keep pushing this. They might as well push a wet noodle across the table. It ain't gonna go nowhere it didn't wanna go.